Language and the rest of cognition: computational, algorithmic and possible neural commonalities

The computational program for theoretical neuroscience proposed by Marr and Poggio [1] calls for a study of biological information processing on several distinct levels of abstraction. At each of these levels — computational (defining the problems and considering possible solutions), algorithmic (specifying the sequence of operations leading to a solution) and implementational — significant progress has been made in the understanding of cognition. In the past three decades, computational principles have been discovered that are common to a wide range of functions in perception (vision, hearing, olfaction) and action (motor control). More recently, these principles have been applied to the analysis of cognitive tasks that require dealing with structured information, such as visual scene understanding and analogical reasoning. Insofar as language relies on cognition-general principles and mechanisms, it may be possible to capitalize on the recent advances in the computational study of cognition by extending its methods to linguistics. Much of the discussion surrounding the integration of linguistics with the other cognitive sciences has traditionally been focusing on arguments for [2] and against [3, 4] the modular [5] status of language. Even if language is a module, however, it may still rely on the same computational principles (in the sense of Marr [1]), and may be supported by the same mechanisms, as the other cognitive functions. To explore this possibility, we need to bring together ideas from several fields. The first, a natural home for the integration project, is cognitive linguistics [6], which consistently produces valuable insights into the psychology of language, yet is little concerned with algorithmic or implementational issues. The second is computational linguistics [7] (including statistical natural language processing [8]), which examines the mathematical nature of language-related tasks and generates important applications, yet pays little attention to behavioral or neurobiological issues. Lastly, there is the Marr-Poggio computational framework [1], which is used across cognition and which spans all the relevant levels of analysis, but has not yet been extended to the study of language. This paper discusses some of the general computational principles that emerge as useful for understanding cognition, focusing on those that are likely to be especially relevant in dealing with structured knowledge. It then brings these principles to bear on a theory of language that is rooted both in cognitive and in computational linguistics, and that views language as an incrementally learnable system of redundant, distributed representations akin to those found by neurobiologists in olfaction, audition and vision.